Steady State Visual Evoked Potentials (SSVEP) based Brain Computer Interface (BCI) provides high throughput in communication. In SSVEP-BCI, typically, higher accuracy can be achieved with a relatively longer response time. It is therefore a research topic to reduce the response time while keeping high accuracy. We propose a new method, temporal alignments enhanced Canonical Correlation Analysis (TACCA), followed by a decision fusion to improve classification accuracy with short response time. TACCA exploits linear correlation with non-linear similarity between steady-state responses and stimulus frequencies. We compare TACCA and three state-of-the-art methods using data from 54-subjects with response time ranging from 0.5 to 4 seconds. The evaluation results show that TACCA yields mean significant accuracy increase of 10-30% in all segment lengths, especially for the shorter time segment. One-way ANOVA tests show high significant differences between single and multiple phases in TACCA performance.